Ship 26 NYC - Opening Keynote

[00:00:08]Hi everyone. It's so great to be here. Since  Ship last year, the world has changed a lot.

[00:00:15]We're writing less code by hand, but more ideas  are coming to life than ever before. In fact, six months ago, less than 3% of Vercel's deployments  were triggered by coding agents or clankers.

[00:00:30]Since then, the number has grown 17 times.  Over half of deployments on Vercel now come from agents. It's wild. But the bigger shift  is what those agents are deploying. Since the beginning of the year, agentic workloads on Vercel  have doubled. Volume on AI gateway grew from two trillion to 20 trillion tokens per month. And  the shift we're seeing is that we're using agents to deploy software that can think. We're  shipping agentic sites and apps, but we're using agents to ship agents. It's a really profound  change. So, I want to go back in time and show you how we got here because our new world is  agentic, but it all started with websites.

[00:01:22]Look at that little G back in Argentina  building simple sites with HTML, CSS, and JavaScript. I was so excited to put my  creations online. And that's what the web is all about. Sharing our ideas with the world. In  fact, we built that first version of Vercel to make that easy for anyone. It was infrastructure  for pages, caching, and global content delivery.

[00:01:50]Today we serve the fastest, most beautiful  websites in the world for companies like the Met, Supreme and Warby Parker. But the web is no longer  static. It's dynamic. Sites became datapowered applications. So Vercel built infrastructure  for servers, APIs, and databases. Now, Zapier, Stripe, and the New York's own TodayTix run apps  at scale on Vercel. For over a decade, we've been building Vercel into a cloud where you can run  everything. And in the last year, we've shipped some huge steps towards that vision. We brought on  some of the finest Python developers in the world to extend our core infrastructure for back-end  frameworks. So now you can run backends like fast API, Flask, ExpressJS and Hono at scale on Vercel.  We built out a complete Agent layer so you can run long running functions on fluid compute. You can  make workflows durable. You can spin up secure sandboxes and you can host your MCP servers.  So you make your app available to any agent.

[00:03:13]And we brought everything else your app needs  right into Vercel from databases to commerce. So you can manage Amazon Aurora Postgres and MySQL  auth providers like Auth0 CMS platforms like sanity and checkout from Shopify all within  Vercel. So this means that you can run any architecture you want on our infrastructure.  For example, OpenAI, a little AI lab you may have heard of. That's right. They run Next.js  frontends on Vercel, Python backends on Vercel, reaching over a gazillion users every  day. And so in addition to front ends, you can host backend only services like REST APIs,  GraphQL, remember GraphQL written in Yes, that's right. Uh written in Typescript or Python. And  so you can deploy workflows and cues that handle longunning asynchronous jobs, billions of messages  every day. You can bring your entire application and Vercel will run all of it. Which reminds me,  people keep asking me when I'm in San Francisco, they stop me on the streets and they say, "What  is the Vercel for backends?" And I tell them, "Vercel [clears throat] is the Vercel for back  ends." Of course, and today we'll prove it. Um, the reality is that all of this software is  becoming autonomous. It's very, very exciting.

[00:04:50]Websites and apps used to respond to user input  with logic, but now they have agents inside that can understand intent and perform autonomous  work. Every new generation of software demands a new generation of infrastructure. So it's natural  that agentic software needs agentic infrastructure and that's exactly what we're building out  at Vercel. Agentic infra has three parts.

[00:05:21]Number one, Vercel is where coding agents deploy  software. So when you ask Claude Code or Codex where to deploy, you get Vercel because Vercel  is built for the way that agents think and work.

[00:05:36]Second, Vercel is where you build and deploy your  own agents. So we give you every tool you need to build and run apps and agents in production  securely at scale in one platform. And third, Vercel itself is automated by agents. So Vercel  runs your apps in production. We handle traffic, observability, traces, anomalies, and so that  data gives our agents the context they need to investigate autonomously and give you PRs, not  just alerts. And so for context, teams using Claude Code deploy to Vercel five times more  frequently than teams who don't. That's because we set the standard for developer experience  and now we're doing the same for the experience of the agents that those developers use. Coding  agents love deployment to Vercel because we give them everything they need. When an agent needs to  verify the work, it needs a live test. So, Vercel gives every deployment a dedicated secure preview  environment. When you ask your agent to ship an experiment, it needs to roll out that experiment  safely. So, Vercel gives every experiment a feature flag, the infrastructure to control and  and roll out that change with confidence with instant roll back and other capabilities. And you  don't want your agent wasting time clicking around a dashboard like it's 1955. That's right. An agent  is most efficient when every part of the platform is available in its own language. And so Vercel  gives the agent a CLI, APIs, MCPs. So this year, as an example, an engineer at Meta opened Claude  Code and asked it to build an internal tool. When it was time to test it, he asked Claude, "Where  do I deploy this internal tool?" and push it to Vercel. A week later, everyone on his team was  deployed to Vercel. Within a month, Vercel was the go-to platform for all of Meta Super Intelligence  Labs. Meta.ai, Meta's frontier AI product, was born on Vercel. And this happened even though Meta  had already spent a bazillion gazillion trillion dollars building their own deployment platform,  building their own infrastructure, but even the most powerful infrastructure in the world doesn't  natively speak the language of agents. Vercel does. And so agentic infrastructure is where you  deploy your websites, apps, and agents. Thank you.

[00:08:45]G just told you that our compute layer powers  everything on Vercel. I want to tell you about it because for years running serverless compute meant  living with tradeoffs. Serverless was supposed to free you from managing servers and it did. But  every one of you who shipped a serverless function knows that it came with a catch. Actually, a  whole list of them depending on the vendor. Your function can only run for so long. There are rules  on how big your function could be. Network access was often wide open. [clears throat] Good luck  getting FFmpeg to work. Trust me, I've been there.

[00:09:30]you get one operating system or in some cases  none. Take it or leave it. There are vendors out there who don't give you real NodeJS or  Python. The limitations on how much you can upload to that function and forget about holding  long connections open. No websockets, no long live streams. Every one of those is a place where the  platform just said no. I'm happy to announce that with Fluid Compute 2.0, that list is gone. You  can run functions for up to 30 minutes. Functions can be up to 5 GB in size. You get private secure  connections to your own backends. You can install any package. You can pick your operating system.  Still, as always, on Vercel, you get, you know, real Node and real Python. Large uploads are  no problem. And there's full websocket support.

[00:10:35][applause] And of course, you keep everything you always  liked about serverless. Automatic security updates and automated scaling. This is 10 years  of innovation building the compute stack of the future. I'm so proud of the hard work our team  has done to give you a world free from trade-offs.

[00:10:57]So, how do we do it? How do we flip every one  of those nos to a yes? Most platforms stitch together a zoo of primitives, bare metal, VMs,  containers, WASM, process isolation, V8 isolates, each with their own rules and trade-offs.  We made one bet instead. We bet everything on micro microVMs and build a platform we call  hive. A microVM gives you a strong no compromise kernel level isolation between tenants. There  are no shortcuts on security. Our bet was that microVMs could model every kind of compute Vercel  runs. Now that journey started with builds a few years ago. Actually today every compute primitive  on Vercel runs on that one unified stack that's built sandboxes and functions. Now a few weeks  ago we brought Docker and OCI images to sandbox.

[00:12:03]But I just mentioned how functions and sandboxes  are now run on one unified stack. So could I run Docker on Vercel functions? Yes, you can.  Today we're bringing Docker images to the whole platform with Vercel container registry and  flexible runtime for fluid compute. [applause] Thank you. You heard that right. You can now run  Docker images on Vercel and truly ship anything you want. Ruby on Rails, PHP, Java, even Net all  work natively on Vercel. And like G said earlier, you can run anything on our infrastructure, front  ends, backends, and backend only services. And now you can run your own containers. But developers  also tell us it's too hard to wire all these place pieces together. And so we fixed it. Today,  I'm excited to announce Vercel services. It's the developer experience you know and love for your  full stack applications. You can now develop your front end and back end together with one  command vc dev and everything spins up locally.

[00:13:14]When you push a commit, you get a preview for  your entire app, not just the front end. Even backend only commits generate a full preview you  can test before you ship. And all services deploy you deploy can talk to each other privately  without ever touching the public internet.

[00:13:33]You can now run all your microservices on Vercel  and everything just works. All right, now let's talk about the elephant in the room. What might  it be? This is all great, but what does it cost?

[00:13:51]You could throw all of this on a $5 VPS, right?  Or as they call it, postinflation the $25 VPS. Uh jokes aside, here's the real model. Fluid compute  uses active CPU pricing. You pay the CPU price only when you're actually computing. So when  your agent is sitting there waiting on an LLM, that is zero CPU cost. This is a compute primitive  and pricing model that was designed for agents.

[00:14:23]Now I know what some of you are thinking, but I  can run my [clears throat] own server cheaper.

[00:14:27]And to be honest, I was kind of thinking that  myself. But can you actually do this? Let's go through this model. Start with us. Fluid is  one flat price. It's about 15 cents per active CPU hour. Again, you only pay for the CPU you  actually burn. A server you rent is the opposite, right? You pay for the whole machine every hour  whether you use it or not. So math in our head, renting only wins if you keep that machine very  busy. Just just to match our price on raw EC2, you'd have to run it at 35% utilization every hour  all year. That's your production systems possible for sure. It's your staging systems and the  developer sandboxes. Everything together must must be above 35% to match fluid. In practice, almost  nobody runs that hot data for real world average utilization kubernetes in 2026 is that companies  on average achieve 8% utilization when considering all those systems. That means you actually use it  about 1 hour in every 12 hours. So that 15 cents CPU hour in fluid compute on EC2 it cost you 58  cents. And if you add Fargate or EKS it's even more. It's the same compute three to four times  the price and you get the privilege of running it yourself. With active CPU pricing you never manage  any of this. You pay for the CPUs and that's it.

[00:16:13]Cool. that summarize Fluid Compute 2.0 Up to  30 minutes of function duration, up to 5 GB of code per function, continued support for VPC  peering, VPCs, VPNs and and static egress IPs, full Docker image support, great DX for Vercel  services and choose your own with continued real support for Node.js, JS, Python, and even up to  100 megabytes of uploads, full websocket support, and active CPU pricing that beats renting servers  in most real world situations. Vercel is now the most capable compute platform for every type of  workload, front ends, backends, and agents. Let me bring up Tomo, our chief product officer, to  show you what you can build and run on top of it.

[00:17:12][music] Hello everyone. Malte, thank you so much. I'm so  happy to be back in New York with all of you. With the latest compute innovations Malte just walked  us through, Vercel is now the ideal platform for deploying any type of full stack application that  includes backends. It's also the ideal platform for building and deploying agents. Agents need  to connect to models, execute complex workflows, and connect to data and apps. And Vercel's agent  stack provides all the tools you need to build and ship truly capable agents. Let me walk you  through these tools now. Starting with how we connect to LLMs using the AI SDK. AI SDK was  initially released 3 years ago last week and has been in active development ever since. It's  a universal toolkit for building AI applications, frameworks, and agents across any model provider.  It's platform agnostic, and it allows you to generate text, images, audio, and more. With  over 16 million npm downloads every week, it's become the standard TypeScript SDK  for accessing any model from any provider.

[00:18:25]And we also re recently released the AI SDK for  Python. AI SDK is being used at massive scale by companies all over the world like Brex. Brex  helps businesses with spend management, expenses, corporate cards, and things like that. They've  built agents that review thousands of transactions at uh uh in order to audit expenses. Those  agents call models, invoke tools, stream results, and parse structured financial data to identify  anomalies. And the AI SDK simplifies all of this with a standard abstraction layer. The best part  is that you can experiment with different models for different parts of your application.  And you can try out new models as soon as they're released, all without needing to make  any changes to your product code. Last week, we released our latest major version, AI SDK 7. It  adds a ton of new capabilities for working across the entire agent life cycle. It adds a unified  API for reasoning, strongly typed tool and runtime context, support for tool approvals, durability,  timeouts, and so much more. But it also enables you to generate videos and build real time voice  apps and agents. Okay, so AI SDK makes building with AI easy, but you still need to manage the  connection to the underlying models and providers.

[00:19:47]That's where our next agent stack primitive comes  in, which is the AI gateway. AI gateway provides a unified interface for accessing AI models, but  it's actually so much more than that. Internally, I've been calling the AI gateway a token delivery  network. And I want to explain what I mean by that. In the early days of the web, there was  something called the hot origin problem. Does anyone remember this? It was actually called the  hot origin problem. Guillermo doesn't agree with me about the the naming here, but this is a real  thing. Popular sites would suddenly receive spikes in traffic from everywhere, but content lived  in only one or a few origin locations. This created overloaded servers, slow downloads,  and a pretty unpredictable user experience.

[00:20:30]The web sort of outgrew the idea that every user  should fetch every asset directly from the origin.

[00:20:37]And so the CDN was born. The CDN became the  internet's performance and reliability fabric, providing intelligent routing, distributed edges,  observability, and so so much more. Similarly, AI use cases have now sort of outgrown the idea  that every token should be fetched directly from the model provider. Tokens have actually become  a production dependency now. We don't just use them during dev. We use them all the time. And  the model labs are becoming the new origins.

[00:21:08]They're powerful, but they're also expensive, rate  limited, and both geographically and operationally variable. As our friends at the model labs  know, this is a really hard problem at scale.

[00:21:24]And so that's why we built the AI gateway. It  routes around failures. It serves tokens reliably through the same global network that Vercel has  run for over a decade. It also simplifies auth, centralizes usage and spend tracking with  granular observability and offers zero data retention. But there's another important problem  that the AI gateway solves, which is model choice.

[00:21:49]Today's agents don't actually use a single model  architecture anymore. They use many models from many providers. Our AI gateway production index  recently showed that teams running agents at scale route across 35 different models. So model  choice isn't just about experimentation anymore.

[00:22:05]It's how you actually run AI in production at  scale. Serhant is one of the most recognized names in real estate. And they use Vercel's AI  gateway to build an AI app that helps their team with property valuation, marketing, and client  communication. They use different models for different tasks based on uh the need for power or  if they're trying to watch cost. Claude performs complex market analysis. GPT is used to create  marketing copy and Gemini generates images. And all of this happens with a single AI gateway API  key. AI gateway is already serving over 1 trillion tokens per day across hundreds of models from  dozens of providers. It's a huge part of what makes Vercel the open platform for AI. Okay, so  Vercel now serves pixels and tokens instantly and reliably, but modern software doesn't really  follow a simple request response model anymore.

[00:23:06]Agents may need to run for minutes, hours, or  even days and across many, many complex tasks.

[00:23:13]So that's where the next tool in the agent  stack comes in, which is the workflow SDK.

[00:23:20]Longrunning workflows and back-end  jobs are nothing new in software, but they're one of the most challenging things  to build. Things can go wrong, and they do. And I'm sure many of you have experienced timeouts,  dropped state, dropped connections, and more.

[00:23:36]Without a primitive for durability, you have to  kind of handstitch together things like retries, state persistence, and this can get really  ugly. This is why we built the workflow SDK.

[00:23:47]We like to say it provides infinite compute  durability. It enables you to build longrunning apps and agents that automatically suspend,  resume, retry, and maintain state with ease.

[00:24:00]DoorDash uses the workflow SDK to run traditional  ETL jobs, keeping data in their app up to date.

[00:24:06]And our customer Flora built an entire AI design  platform on the workflow SDK to help designers generate visual content at scale. Inside their  platform, agents fan out across 50 different image models to generate visual directions from  a single creative brief. Workflow SDK checkpoints every step of every agent and pauses when the job  needs human input. And because every failure is automatically retried, the designer never needs to  start over. The whole idea behind agents is that they can solve complex tasks across multi-step  workflows. And what we found is that one of the primary ways agents want to accomplish tasks  these days is by writing and executing code.

[00:24:51]That's where our next agent stack primitive comes  in Vercel sandbox. LLMs have gotten really good at producing working code to accomplish a task. But  that code is still untrusted and it can't be run inside the same environment that has access  to our production systems. We need a special production grade environment that's designed  for executing untrusted code in a secure and isolated way. Vercel is actually no stranger to  this problem of untrusted code execution because of preview deployments and builds. We host over a  billion production grade preview deployments and run over 6 million builds every single day. Every  one of those builds happens inside an isolated microVM compute environment. With Vercel Sandbox,  that same isolated compute environment is now available to you and your agents. It's the safest  way to write code, run code you didn't write, and it's perfect for agents. Each sandbox is a  fully functional computer with a file system, security boundary, and even full Docker  support as Malte just mentioned. And of course, it's all built on top of fluid compute. Vercel  sandbox is already being used in production at scale by the best companies in the world,  including notion. Millions of teams use Notion as their AI workspace, capturing knowledge,  answering questions, and moving projects forward.

[00:26:13]And now you can extend Notion with custom agents.  This enables you to do things like sync CRM data, turn Slack threads into content, and connect to  external workflows. But custom agents need code, and that code has to run safely, which is why  Notion runs custom agents on Vercel sandbox. Since each agent gets its own general purpose compute  environment, everything outside the sandbox stays protected. Okay, so now we can execute code  securely inside long running workflows. But in order to do anything useful, our agents need to  access data and tools. And that's where the next layer of the agent stack comes in. Starting with  Vercel connect. Vercel connect is a brand new building block that allows you, your agents, to  securely connect to all the data and systems they need. With Connect, your agents can access tools  like Slack, GitHub, Salesforce, Snowflake, and so many more. And you can even create your own custom  connectors with OAuth and API keys. But here's the most important part. Instead of longived  secrets, every connection uses a short-lived, minimally scoped access token. So an agent never  has standing access to your systems and it never touches data it shouldn't. This short-lived  access token point is really really important and minimal minimal scoped access. Zapier makes  it easy for people to build AI automations and workflows. They use Vercel connect to give their  apps private IP stable network access to internal AWS resources like RDS, EKS and more via private  link and VPC peering. That means their apps can use real live business data, not just whatever is  reachable via a public API. With Vercel Connect, your agents can securely connect to your full  range of internal systems. Your CRM, your ERP, your HRIS, data warehouses, and more. But one  of the most important connections your agents need is to the tools we use to interact with them.  And that's where the next primitive in the agent stack comes in, which is the chat SDK. Chat SDK  provides an elegant abstraction layer enabling you to interact with agents across dozens of apps.  The most powerful agents we've built at Vercel are effectively our co-workers and we interact with  them where work is happening as it's happening.

[00:28:37]With just a single line of code, chat SDK enables  your users to interact with agents across tools like Slack, Teams, Google Chat, Discord, GitHub,  Linear, Telegram, WhatsApp, and so many more. Nano Claw helps companies run AI agents, and they  built their platform on top of the chat SDK.

[00:28:56]It's a single codebase, but it delivers agents  across 15 messaging apps. So that's the Vercel agent stack. It's a singular set of endto-end  capabilities covering everything needed to build and ship agents to production. We've taken  everything we've learned building agents over the past few years and turned those learnings into  best-in-class primitives that work at Vercel scale. We love this stack and our customers do  too. It's a powerful stack that fills a real gap in the ecosystem, but there's still quite a bit of  complexity. Each of these primitives still needs to be wired up into a single cohesive agent. What  if they didn't? At Vercel, we don't want to just build the world's most powerful agents or enable  you to build the world's most powerful agents. We want those agents to have craft. We want building  them to be simple. And that means we also have to provide you with the fastest and easiest way to  stitch all of those primitives together. To make that possible, we built the newest member of the  agent stack family, Eve. Thank you. [applause] Eve is our platform agnostic framework for  building complete endtoend production agents.

[00:30:11]It's opinionated based on everything we've learned  over the past two years, but critically it's also completely open- source and modular. It's built  to work seamlessly with uh our infrastructure with high cohesion between Vercel and Eve, but  it's completely customizable, so you can make it your own and run it anywhere you'd like. With  Eve, every agent is just a directory laid out in a truly intuitive way. We like to call it  the Next.js for agents. I have a lot more to share about Eve, but first I want to welcome  Char to the stage to walk us through a demo.

[00:30:49][music] Thanks, Tomo. Over the last year, we've taken  everything we've learned about building agents and we packaged it into Eve. Let me show you how  it works. I'll get started with a single command in my terminal. This command will scaffold  an agent directory, install dependencies, and start an interactive chat session with the  agent. Next, I'll configure a model provider.

[00:31:18]I'll select Vercel AI gateway, select the  Vercel team, links to a project in that team, and we're done. Now, let's  test the agent. Who are you?

[00:31:37]Okay, that was fast. I just built a fully  functional agent in less than a minute. Now, this agent runs with just two files.  Agent.ts, which configures the agents model, and instruction.md that defines the agent's  identity. [clears throat] And that simplicity is what makes it so easy to build an agent  with Eve. Now let's build a real use case, a go to market sales agent that processes call  transcripts and updates Salesforce and linear.

[00:32:07]Let's build that agent with Eve. And first I'll  give the agent an identity. [clears throat] In the instructions markdown, I'll describe  a go to market assistant that can create linear issues and update Salesforce  opportunities based on call transcripts.

[00:32:23]And just like that, the  agents have a updated charter.

[00:32:29]Next, the agent needs to create issues. So,  I'll connect it to Linear. And connections are how Eve plugs into external MCP and Open API  servers, and they go in the connections folder.

[00:32:46]I'll define a MCP client connection and  use the linear connector in my Vercel.

[00:32:52]And I can limit the linear  MCP tools the agent can use.

[00:32:58]I also want the agent to know how to create linear issues. So I'll add a skill and  skills go to the skills folder.

[00:33:08]This skill will tell the agent it should look for  feature requests in the transcript and separate linear issues by topic. Now let's test the  agent. It should be able to create linear issues.

[00:33:26]Great issue was created and we  should also see it in linear.

[00:33:36]Next, the agent needs to update Salesforce  opportunities. We want to use the Salesforce CLI to make updates and the commands to be  executed safely in isolation. I'll configure a sandbox.ts. This sets up a private VM for the  agent to work in. And the sandbox is only allowed to talk to Salesforce and nothing else. I'll give  the agent a Salesforce tool in the tools folder.

[00:34:04]This will run the Salesforce  CLI inside that sandbox. And Salesforce updates need approval from  someone on the team. So I'll also give it human in the loop. So it always  requires an approval from a human.

[00:34:16]And lastly, the tool execution function that will  only update Salesforce opportunity. Now, let's test the Salesforce step. I'll ask it to update  Salesforce opportunity. That's human in the loop.

[00:34:33]And it's updated. And if you go to Salesforce  refresh, we should see the updated status.

[00:34:41]Great. Eve also comes with eval support out  of the box. So I can run the agent against real real sessions and grade the results.  Eve discovers eval under the evals folder.

[00:35:00]I'll set a config and specify a model to  judge the evals. Then create a transcript with two different requests and check  that the agent opened two issues, split them by topic, and didn't  touch Salesforce on its own.

[00:35:16]Now let's run the eval.

[00:35:25]Great. It passed. Now a prompt change can  quietly break the agent. We are now ready to make the agent available in Slack  for the sales team. So back to the TUI.

[00:35:38]I'll run slash channels.

[00:35:43]Add Slack. Yes, I want a Slackbot.

[00:35:50]Create a Slack connector. This will  install the Slackbot in the workspace and deploy to Vercel.

[00:36:04]Let's test it end to end. I'll give it an  example transcript in Slack. Tag the agent.

[00:36:16]It'll ask me to approve the change.

[00:36:21]And that's it. We built a production ready agent  that our team can use. But we're not done yet.

[00:36:27]This agent is also fully observable. So let's  take a look at that conversation we just had with it. In the Vercel dashboard or click  on observability agent runs and here's the conversation. I can see the entire conversation  history. The token usage, the length of run, the inputs, outputs and the agent reasoning. I  can even drill down into each of the tool call the agent made. And lastly, Eve is open source.  You own your code. You can swap the primitives.

[00:37:00]And Vercel is the fastest way to deploy  it, but it's not the only way. And that's it. We just build a production ready agent  in five minutes. Back [applause] to you.

[00:37:13][applause] We launched Eve two weeks ago at Ship in London  and it's been incredible to see what everyone has built already. Over 1,500 teams have deployed  over 40,000 agents on Eve and it's got over 250,000 npm downloads in the last week. But  developers aren't just experimenting with Eve, they're running agents in production at scale.  Clay built a data analyst agent on Eve. Solomon built an incident response agent on Eve. And Bask  Health built an agent that lets non-technical users design and preview product features.  Here's a quote from the team at Bask Health that I'm really proud of. We've gone through a  few different iterations over the last 6 months, but Eve has been the most batteriesinccluded  experience. I can't believe how quickly we got up and running. This was exactly our goal.  Now, we've built a ton of agents ourselves, as we've mentioned a few times, but one that I'm  particularly proud of, we gave the Vercel name.

[00:38:21]We just call it Vercel agent. I like to describe  it as an intelligence layer for Vercel because it can help you do almost anything on the platform.  We built a beautiful web UI so you can chat with Vercel agent directly from the dashboard and ask  it almost anything. For example, you can ask it to fix the 500s that are showing up in your logs and  it'll perform an investigation. You can ask it to find accessibility issues in your project and  it'll run a review and open a pull request. Or you can simply ask it to fix your build and it'll  read the deployment logs, find the failing config, validate the fix in a sandbox, and redeploy with  your approval. But what I love most is that the Vercel agent is not just another chatbot. It sees  your app running in production and proactively monitors your infrastructure on your behalf. Let  me give you an example. Let's say you get an alert about a partial outage. When you log into the  dashboard and click on the alert, Vercel agent will have already run an investigation on the  anomaly. Here we can see it found that the API key object is undefined at runtime and errors appeared  after the last deployment a few minutes ago. It recommended an instant roll back to the previous  release, which when in doubt is always the right action to take. Um, so this one's an easy one to  approve. With permission, Vercel agent rolled back to the last production deployment before the 500s  showed up. And after the roll back was successful, our app is back online and Vercel agent can kick  off work on a proper fix for the error. Now, the first question every CTO asks is, is it  safe to let the agent do that? And this is the right question because most agents inherit their  users permissions. They run as you do and can do everything you can do. So, a single bad prompt can  have extremely wide blast radius. But Vercel agent has a first of its kind permission model. It has  its own identity and it's read only by default.

[00:40:23]First, Vercel agent presents you with a plan,  not an unknown number of approvals. And instead of asking, can I do this? Can I do this? You  approve the entire plan upfront. So you know what the outcome is going to be before Vercel  agent does any work. Then it's granted permission scoped exactly to that plan and nothing more.  Every agent runs in a sandbox before it touches production and anything that changes your  production apps for your approval. So it never has more access than the task needs and it never  stops to ask you again. Vercel agent is just one of the many agents that we've built. Uh and I'm  excited to welcome Jeanne, our COO, to the stage to tell you about all of the agents we've built  to transform our go to market motion. Thank you.

[00:41:19]Thanks Tomo. Vercel agent shows you what what's  possible when you build agents the right way. And every company in this room is going to build  an agent just like it. There are two types of people in this room hearing that. Some of you are  sitting there thinking, "Let's go. I am shipping tonight." And then there are the CIOS and the  CTO's in the room who are thinking, "Oh no."

[00:41:44]Because you can already feel what's coming. Shadow  agents writing to systems with no audit trail.

[00:41:50]AI bot user closing your tickets spend you can't  explain. Both of you are right. Building agents is easier than you think and way harder than  you think. I'm going to tell you about that tension and what we've learned at Vercel from  living it. Drew Bredvick, who works for me, is head of go to market engineering. In June of  2025, he had the sexiest job in the world. His mandate, build the agents that would transform how  Vercel goes to market. It worked. A year later, agents are part of our daily workflow running at  scale across our entire go to market organization.

[00:42:31]You did a great job, Drew. So today, I'm handing  you a pager. Let me tell you why. What we learned was that agents are free. As in free puppies.  Everyone loves puppies, but they pee on your floor. They eat your furniture, and you can't go  on vacation. Agents are free because anyone can prompt Claude. But agents are also software, and  we all know that software is never done. Someone has to maintain them, update models, and build  new features. Building hundreds of agents taught us hard lessons. First, we saw the same problem  solved over and over. Multiple agents needed to connect to the same internal systems. Each team  built their own integrations from Slack. Second, each agent was reading from different knowledge  bases. Our team would ask the same question and get different answers. Third, we had no  visibility. No one knew how many agents existed, who built them, or what data they touched.  Fourth, that lack of data of visibility also meant adoption chaos. I'm on Slack, I type at 500  agents pop up. I don't know what any of them do.

[00:43:49]And last, we also learned that chat isn't all that  you need. We started with the idea that Slack was the universal interface, and that was wrong. The  agents that actually got used also had frontends for permissions, for visualizing data, for  workflows, and for keeping humans in the loop. We learned that our agents had to work on day one and  day 100. And I'm happy to tell you that they do.

[00:44:18]We run over 100 agents in production at Vercel and  they're part of how we operate every single day.

[00:44:26]I want to tell you about 10 of the most important  ones and the order matters. We started with the most obvious use cases and worked our way towards  the agents that changed our internal processes and transformed how we operate the goto market team.  First, Vert.Ex is our customer support agent. It resolves over 91% of support tickets across the  help center, Slack, and Docs chat. Second, deal one is our deal intelligence agent. It listens to  every sales call, coaches our sales reps in Slack, and runs a post-mortem on every lost deal. The  Deal One MCP has been called 17,000 times in the last month. Next, draft zero is our content agent.  It writes the first draft of every blog post, change log, and customer story we publish. A Z is  our AEO agent. It tracks how Vercel's brand and content shows up across AI search. Every day it  runs hundreds of prompts across dozens of coding agents. Ravoa is our Salesforce update agent.  It publishes crit critical record changes to Salesforce with a human in the loop. It saves nine  hours of time for our RevOps team every single day. [clears throat] And Penny is our finance and  ops agent. It has access to our billing platform, payment provider, and monitoring systems. It saves  our finance and on call engineering teams hours of answering billing questions. And then there's  V. V is our routing agent. It routes requests to all of our other internal agents. Remember that  adoption problem? We still have 100 agents, but V is a front door to all of them. ask V a question  and it picks the right agent for the job. The last three I'm going to show you are the ones I want  to spend real time on because every one of you is going to want to build them. First, D0, our data  agent. D0 gives the entire company 247 ondemand analytics and data science. Anyone at Vercel,  engineers, AEs, finance, support can run analytics on data from our warehouse without filing a ticket  or waiting on the data team. Users can ask simple questions like how many leads we got from this  campaign. And D0 writes and runs SQL. But D0 is also a data scientist. If users need statistical  analysis done, it spins up a sandbox and runs Python to generate reports. D0 is the most  used internal tool at Vercel. It handles 30,000 questions per month and it's safe at scale. D0  doesn't run in god mode. Every query is scoped to the user's permissions. If you can't see a table  in Snowflake, D0 can't show it to you either.

[00:47:28]Under the hood, D0 needed a semantic layer.  And even though you can ask questions in Slack, it needed a UI so people could explore charts and  drill into data. Next, Athena, our sales cockpit.

[00:47:43]Salesforce announced headless. We've been running  on it for months. Athena picks accounts, plans outreach, tracks signals, and runs the weekly  motion for every AE at Vercel. Shortly after we went live, pipeline coverage nearly doubled.  Every AE uses it every single day. Under the hood, Athena needed the same semantic layer D0 needed,  plus durable workflows and secure connections, plus a UI because GTM agents are more effective  with pixels and buttons than just a Slack channel.

[00:48:18]Third, our lead agent, an autonomous SDR. We  trained lead agent on the playbook of our best SDR, and now it runs that playbook 247. You may  have seen the headline last year. We took 10 SDRs down to one. That was lead agent. We redeployed  those nine reps into bigger roles and raised our quota. We've seen a 32x ROI and it costs $5,000 a  year to run. It performs at the 90th percentile of our reps and one engineer maintains it part-time.  Under the hood, lead agent runs on the stack Tomo just showed you. AI SDK, Workflows SDK, Chat SDK,  and it's open source, so you can go build your own today. Building these agents was easy because  we use Next.js and our agent framework, Eve, and running them was never a problem because they  run on Vercel. Remember when I said that building agents is way harder than you think? The hard part  is everything around them. who can access them, how they authenticate, what data they can touch,  and providing all of this to your security team.

[00:49:32]So, we spent the last year building the  platform that makes that easy. And today, we're making it available to you. I'm excited to  announce Vercel for enterprise apps and agents.

[00:49:46]It's the Vercel developer experience you  love for everyone at your company with identity and access builtin and option to run it  in your own AWS tenant. Like all of our products, it's portable because it's framework and model  agnostic. We built it for ourselves first and now it's the platform you can build on. Let me show  you the most important pieces. I said earlier that anyone can prompt Claude. The reality is that  your employees are already doing this whether you know it or not. That's called shadow IT and  AI has already caused major data breaches in the enterprise. So even if you control who builds AI,  you still have to limit access to the apps and agents they build. That's where Vercel Passport  comes in. Whether you're using v0, Claude Code, Codex, everything your team builds stays internal,  governed by the identity pro provider you already run like Okta or Microsoft Entra. And it goes  deeper than the front door. Passport issues a token tied to each user's identity and carries it  through to the backends and systems your apps and ages agents connect to. At Vercel, we use v0  to build data apps and reports powered by our native Snowflake integration. This allows us  to build apps that have direct access to our Snowflake data. But Passport doesn't just protect  access to the application, it also limits the data users can see in the application based  on permissions provided by our IDP. I'll give you another example from our customer Mercor.  Their security team built an internal agent that understands their entire codebase. You can  ask it questions like which of our public routes uh connect to a database and it will pull  answers across every repository. It'll even kick off fixes. But what makes it interesting is  not just people asking the agent questions. Other agents query it too. So they had a hard problem.  They wanted employees and agents to be able to use it and they wanted to control access but they  didn't want to hand and manage authentication for it. They implemented passport and connect in  under a week and now the agent runs on Vercel behind Okta the IDP they already use. The right  people and agents get in scoped automatically with no homegrown auth. Passport controls who gets  into your apps. But when everyone is shipping, you also need the ability to look across  deployments and see if there are security risks.

[00:52:29]That's why we built the security dashboard. It's  one place to see every secure see the security posture of every account and project on Vercel. It  flags misconfigurations that can lead to breaches, shows you members without MFA, alerts you  about secrets shared across projects, and identified long lived credentials when short ones  will do. The best part is that it actually shows you how to fix each problem. I'm an excited to  tell you that today all of those deployments can run in your own AWS tenant. You heard that  right. You can run Vercel functions in your own AWS account. Whatever you build with agents stays  inside your cloud. Everything I've shown you today runs on Vercel's core platform. It's the same  platform the world's best teams have trusted for over a decade for our developer experience,  previews, firewall, end-to-end observability, and features like instant roll back. These new  capabilities aren't bolted on. They're part of the Vercel you already trust. The teams who ship  fast, securely at scale are the teams who will win. We built enterprise apps and agents so  that can be you. G, back to you to wrap up.

[00:53:52][music] Thanks, Jeanne. Today we showed you that Vercel is  a platform where you can build and run everything any kind of website, app or agent. Your backend  frameworks can run at scale. With Vercel services, you can develop, preview and ship backends  and front ends together. Fluid compute, as Malte showed us, lets you run any type of  workload from instant serverless functions to back-end agent jobs that take hours, days,  months to complete. And you can even bring your own container. With Eve, your team can build  a production agent in minutes. Vercel Passport makes sure your internal agents and apps stay  internal behind your IDP. And you can run all of this in the security boundary of your AWS account.  It's beautiful. And you can start today. Ask your coding agent to install the Vercel plug-in. Then  you can build anything a website, an app, an agent and ship it at global scale. And Vercel agent of  course will keep an eye on production. This is Agentic infrastructure. We can't wait to see what  you all ship next. Thank you very much. [applause]